NovaPunk LLC
← All services

01 / AI Agents

Custom AI Agents

From customer signals to personalized action.

NovaPunk builds AI agents around a defined business workflow: understanding customer signals, selecting the next action, generating relevant content, and connecting the result to operational systems. For marketing teams, this can bring audience targeting, product recommendations, personalized offers, delivery, and campaign visibility into one application.

Discuss this service

Built for your requirements.

The problem

Sending the same message to every customer ignores purchase history and product preferences. The example connects those signals to audience selection, relevant offers, and personalized email content.

What we can build with you

  • Business requirements, campaign rules, and an end-to-end workflow design
  • Customer features, targeting models, and explainability views
  • Product and brand knowledge retrieval connected to content generation
  • Personalized offers and channel-specific content
  • Delivery integration, activity history, and campaign dashboards

Turn a campaign brief into an operating workflow.

The service starts with a business objective and the rules that define a valid action. Data, prediction, generation, and delivery are designed together so the application produces a usable result.

Customer understanding

Connect purchase history, product categories, and customer attributes. Derive recency, frequency, monetary value, and other features that support a targeting decision.

Targeting you can inspect

Build a model for the available data and evaluate it before use. In the Walsh example, a Random Forest ranks customers and feature importance helps inspect the signals behind the model.

Context-aware offers

Combine customer preferences, catalog details, discount rules, and retrieved brand information. Generate an offer with relevant context rather than relying on a single generic campaign template.

Execution and visibility

Connect generated content to an approved delivery channel and record the output. A dashboard shows selected audiences, recommendations, model scores, and campaign activity.

A Mother's Day campaign, end to end.

The research example illustrates how the service components fit together. These steps describe the supplied project, rather than a promise of identical results for every business.

1. Prepare the customer signals

Dataform creates customer features from synthetic orders: recency, order frequency, total spend, average and maximum order value, category breadth, quantity, discount behavior, and favorite brand.

2. Select a relevant audience

Vertex AI Colab Enterprise trains a Random Forest and scores 200 customers. The 30 highest-scoring customers are selected. The report defines Premium at 0.85 or above with a 20% discount, and Very High from 0.75 to below 0.85 with 15%.

3. Retrieve and generate

A corpus of 500 synthetic brand reviews is embedded and indexed in Pinecone. LangChain retrieves relevant passages and combines them with the customer profile and product information for Gemini offer generation.

4. Present, deliver, and record

Looker Studio connects to BigQuery for campaign overview and offer-detail views. Gemini creates a subject line and HTML email; the notebook demonstrates SendGrid submission. Separate email_log and email_content tables preserve status and generated content.

One workflow, different customer touchpoints.

Email is the implemented demonstration. The report also identifies ways to reuse offer outputs in other channels; these extensions must be scoped and integrated separately.

Email and campaign operations

Personalized subject, greeting, product description, offer body, and call to action, supported by an operational view of the generated offers.

Additional channels

SMS, outbound call scripts, direct mail, web promotions, and push notifications are possible extensions of the offer data. They are not presented here as completed integrations.

AI Agent - complete system architecture.

The agent connects four responsibilities: ingesting business data, preparing a reusable analytical view, selecting and generating the next action, and delivering outputs with operational visibility. The Walsh Retail example shows how these responsibilities map to GCP and the AI components.

Operational data → analytical features → prediction + retrieved knowledge → personalized action

Workflow orchestration coordinates the data preparation; campaign records connect model outputs, generated content, and reporting.

Layer 01

Business data and ingestion

Customer profiles, orders, product/SKU records, and related entities form the operational source. Initial migration establishes Cloud SQL tables; Datastream captures subsequent database changes for the BigQuery replica.

Cloud SQL · Datastream CDC · BigQuery source replica

Layer 02

Analytical features and knowledge

Dataform joins the source records into customer features such as recency, frequency, spend, and preferences. A separate review corpus is embedded and indexed to supply contextual knowledge during generation.

Dataform · mart.customer_features · Embeddings · Pinecone

Layer 03

Prediction and offer generation

A Random Forest scores customers and writes the selected targets. LangChain retrieves relevant knowledge, while Gemini combines that context with products, customer preferences, and segment-based discount rules to generate offers.

Vertex AI · mart.ml_prediction · LangChain RAG · Gemini · campaign.offer

Layer 04

Delivery, reporting, and records

Gemini turns an offer into a personalized subject and HTML email. SendGrid receives the message. Looker Studio exposes target and offer information; campaign tables preserve submission/delivery status and generated content.

Gemini · SendGrid · Looker Studio · email_log · email_content

Walsh Retail infrastructure, data pipeline, intelligence layer and reporting architecture
Original system diagram from the Walsh Retail report, Figure 39. Its historical labels include Kaggle brand reviews and Tableau/ROI analysis; the later implementation text describes synthetic reviews, Pinecone and Looker Studio. No measured campaign ROI is presented here. Select the diagram to enlarge it.

Automation Flow - from data to delivery.

Continuous ingestion and scheduled feature preparation supply the campaign workflow. Prediction, generation, delivery, and reporting follow the prepared data.

  1. Capture operational changes

    Datastream reads the source binary log and propagates changes into BigQuery. This ingestion is continuous and distinct from the scheduled transformation run.

    Output: An updated analytical replica of the operational source.

  2. Compile and refresh customer features

    Cloud Scheduler triggers Cloud Workflows daily at 01:00 UTC. Workflows creates a Dataform compilation result and submits a workflow invocation for the SQLX transformations. The documented invocation return is not, by itself, verification that all downstream campaign stages completed.

    Output: mart.customer_features and a traceable transformation invocation.

  3. Score and select campaign targets

    The Vertex AI Colab workflow reads the prepared features, trains/evaluates a Random Forest, and scores the customer population. The example selects 30 targets from 200 customers and records scores, segments, model version, and batch information.

    Output: mart.ml_prediction with the selected audience.

  4. Retrieve knowledge and create an offer

    Customer preferences guide retrieval of relevant review passages from Pinecone through LangChain. Gemini receives the retrieved context, recommended product, and campaign rules, then produces an individualized offer.

    Output: campaign.offer with product recommendation, discount, and generated text.

  5. Generate the personalized email

    Join the stored offer to customer and product information. Gemini returns the subject and HTML body, including greeting, rewritten product description, offer text, and call to action.

    Output: Channel-ready email content for each selected customer.

  6. Submit and preserve the activity

    The notebook submits the generated message through SendGrid. HTTP 202 records provider acceptance; an inbox example separately demonstrates receipt. Status history and full generated content are stored in separate campaign tables.

    Output: Submission/status records and archived email content.

  7. Review the campaign

    Looker Studio reads the prediction and offer tables to show audience size, model scores, product/category distribution, and offer details. Open/click tracking, A/B testing, and feedback into retraining are identified as later integration work.

    Output: Campaign visibility for the business team, with a defined path for future feedback.

Campaign offer inputs flow through Gemini to subject, offer body, product description and CTA, then SendGrid delivery
Email generation and delivery diagram from the report, Figure 46. The demonstration runs in Colab; the Cloud Run Job shown in this diagram is the proposed production execution path. Select the diagram to enlarge it.
Demonstrated workflow and production automation

The reference project demonstrates CDC, scheduled Dataform invocation, customer targeting, generated offers, notebook-based email submission, and stored campaign records. Scheduling the complete ML-to-delivery chain, monitoring completion, retries, duplicate-send prevention, and response-driven model feedback are production integration work.

Internal research / prototype

Walsh Retail marketing automation

Walsh Retail is a fictional retailer developed for an internal Mother's Day marketing automation project. The supplied report follows the complete engineering path from Cloud SQL and BigQuery through Random Forest customer selection, RAG-assisted offer generation, Looker Studio reporting, and an email demonstration using Gemini and SendGrid.

Static demonstration page. Figures and examples replace the video presentation for now; no live agent, model, or simulator runs on this page.

Inputs and data

  • 200 synthetic customers and their purchase histories
  • Product catalog and category preferences
  • RFM features derived from transactions
  • 500 synthetic brand reviews used for retrieval

Example scope

The supplied report records 200 customers scored, 30 selected targets, and 30 personalized offers. It also includes a rendered generated email, a SendGrid HTTP 202 response, an inbox example, and delivery/content history tables. HTTP 202 confirms provider acceptance; the inbox screenshot is a separate delivery example. No measured uplift in conversion or revenue is reported.

Campaign workspace

Illustrative layout / synthetic examples
Example records show the intended workflow; they are not exported execution logs.
CustomerSegmentProduct contextNext action
Sample customer APremiumPreferred category + retrieved reviewReview personalized offer
Sample customer BVery HighPurchase history + catalogReview generated email
Sample customer COtherCampaign selection rulesExcluded from target audience
Offer → content → delivery → activity log

Client-specific approval steps can be added before execution.

Walsh Retail Mother's Day Looker Studio dashboard with customer counts, prediction score, order categories and offer distribution
Actual dashboard figure from the supplied Walsh Retail report, Figure 44. The data is synthetic; 0.986 is the average model score, not a measured conversion rate. Select the figure to view it at full size.
Walsh Retail offer-detail dashboard showing customer IDs, favorite brands, generated offers, discounts and recommended products
Actual offer-detail view from the supplied report, Figure 45. This shows prototype output records rather than a live dashboard. Select the figure to view it at full size.

What the study reports.

200Synthetic customers
30Target customers in the write-up
30Personalized offers reported

Metrics are reproduced from the supplied research or original project account. They have not been independently rerun for this website.

Boundaries and next work

  • The customer and order histories are synthetic and Walsh Retail is fictional; the report also uses a public product catalog.
  • Model scores and synthetic demonstrations do not establish real campaign conversion or revenue uplift.
  • The generation and delivery demonstration runs in a notebook; full scheduling and Cloud Run Job execution are described as production integration work.
  • Real vendor knowledge, A/B testing, open/click analytics, approval policies, and model feedback remain implementation-specific work.
  • The website shows stored examples only and sends no campaign messages.

Explore the engineering.

Based on Walsh Retail Marketing Automation - Sungwon Son, April-May 2026, with attribution to the original series at isupernova.io.

What would you like to build?

Tell us about your workflow, available data, and the decision or task you want to improve.

Contact NovaPunk